## TL;DR
A triage agent misroutes when its confidence is low but it routes anyway, or when queue definitions overlap so the "right" queue is ambiguous. Require a confidence threshold below which the agent asks or defaults to a human-reviewed queue, detect bounce loops where tickets cycle between queues, and treat repeated misroutes as a labeling problem: the queue definitions need sharpening, not the model.

## The query

```text
support triage agent routing tickets to the wrong queue: loop detection
```

## Use this when

- AI triage sends tickets to the wrong queue
- Tickets bounce between queues
- Misroute rates climb after changes
- You are operating an AI triage agent

## Not for

- Training the triage classifier
- Defining queue responsibilities
- Training human triagers
- Ticket assignment within a queue

## Steps

### 1. Measure the misroute rate per queue pair

Track which queue pairs the agent confuses: billing to tech support, refunds to general. The confusion matrix shows whether the problem is one ambiguous boundary or general poor accuracy. Fix the worst pair first.

Expected output: a confusion matrix of predicted versus correct queues.

### 2. Add a confidence threshold with a fallback queue

Below the threshold, the agent routes to a human-reviewed triage queue instead of guessing. A wrong guess costs more than a review: it delays the ticket and annoys two teams. Set the threshold from the measured accuracy curve.

Expected output: low-confidence tickets land in review, not in wrong queues.

### 3. Detect bounce loops explicitly

Flag tickets that change queues more than twice. A bounce loop means two queues each think the ticket belongs to the other, which is a definition problem. Alert on bounces and route loopers to a human immediately.

Expected output: bounce detection with human escalation on loops.

### 4. Sharpen the queue definitions the agent uses

Rewrite the queue descriptions with examples and explicit boundaries: what belongs, what does not, and the tie-break rule. The agent routes by these definitions; vague definitions produce vague routing.

Expected output: queue definitions with examples and tie-breaks.

### 5. Review misroutes weekly and feed corrections back

Sample misrouted tickets weekly, label the correct queue, and feed them back as training or prompt examples. Misroute review is the maintenance loop that keeps accuracy from decaying.

Expected output: a weekly review turning misroutes into corrections.

## Template: the checklist

```text
Triage misrouting:
[ ] Misroute rate measured per queue pair (confusion matrix)
[ ] Confidence threshold with fallback to human review
[ ] Bounce-loop detection with human escalation
[ ] Queue definitions sharpened with examples and tie-breaks
[ ] Weekly misroute review feeding corrections back
```

## Variant phrasings

### ai triage wrong queue

Steps 1 and 4. Measure the confusion, sharpen definitions.

### tickets bouncing between queues

Step 3. Detect the loop, escalate to a human.

### triage agent low confidence routing

Step 2. Threshold plus fallback queue.

## Why it happens

Classifiers route by similarity to training examples, and queue boundaries in real support orgs are fuzzy: a billing question about a technical failure genuinely belongs to both. Without a confidence threshold the agent is forced to guess on ambiguous tickets, and without loop detection the guesses bounce. The fix is decision policy around the model, not a better model.

## Edge cases

- New queues: a new queue has no training history. Start it in human-review mode until it accumulates examples.
- Seasonal shifts: product launches and incidents shift ticket mix. Re-check thresholds after big launches.
- Multi-issue tickets: one ticket, two queues. Define whether to split, pick primary, or dual-route.
- The fallback queue overflowing: if most tickets land in review, the threshold is too high or the model needs retraining.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_nHa80oaB69EKCfVe2z8z6g